Deep learning approach to cervical segmentation from routine CT images

Recently, techniques for 3D medical image segmentation have become increas ingly sophisticated. Different types of Unet-based segmentation networks per form very well in tasks where the liver, blood vessels or brain are the seg mentation targets. Yet, compared to these holistic segmentation targets, another equally important class of human tissues, the bone, has a more pronounced se quential character. Traditional 3D methods are unable to learn the exact position of bone joints due to the limitation of local similarity and numbers of param eters. To address the sequential and location-dependent nature of bone segmen tation information, the dissertation proposes a pre-training model for location information, and combines 2.5D image input and labeled multi-channel coding to semantically segment 3D cervical spine CT images. This method can accu rately extract the information of bone joint location of slices and incorporate it into image features in the form of channels, which can improve the accuracy of multiple bone joint segmentation and classification. The experimental results show that this approach achieves better segmentation results in the test set.Master of Science (Computer Control and Automation